Anthropic has just published a new paper with a straightforward title: **Verbalizable Representations Form a Global Workspace in Language Models**.\n\nPut simply, the research team found that modern large language models have a special internal representational space called **J-space**, short for Jacobian space. Only a small fraction of all the model's computation happens in this space, and the content it holds can be "verbalized" by the model, actively manipulated, and used for multi-step reasoning. Nearly all of the model's remaining processing is automated and unconscious.\n\nDoes that sound like the global workspace theory of the human brain? Yes, that's exactly the analogy Anthropic draws.\n\n\n\n## What Is J-space?\n\nJ-space was discovered using a new technique called **Jacobian Lens**. In simple terms, it calculates the first-order impact of intermediate layer activations on the final output tokens to identify vector directions that are "ready to be verbalized".\n\nThe concepts corresponding to these directions are activated internally even if the model never ends up outputting them. For example:\n\n- When asked "What color is the fourth planet from the Sun?", the model first activates "Mars" internally, then activates "red", before finally outputting "red".\n- When asked to write a line of poetry, it will pre-activate the rhyming word needed for the second line right at the end of the first line.\n- When given a protein sequence, it reads out "protein", "fluorescent", and "green" in its intermediate layers.\n\n\n\n## Five Key Properties\n\nThe paper uses five experiments to prove that J-space behaves very similarly to human "access consciousness":\n\n1. **Reportability**: The model can verbalize content stored in J-space. If you inject a concept such as "elephant", the model will report it in its response.\n2. **Directional Modulation**: You can ask the model to "write a sentence while calculating 3² - 2 in your head", and Jacobian Lens will confirm that it is indeed performing this calculation.\n\n3. **Internal Reasoning**: Intermediate steps of multi-step reasoning (for example, first calculating 21, then 42, then 49) all appear in J-space, and swapping these intermediate concepts changes the final answer.\n\n4. **Flexible Generalization**: A single J-space vector can be read by multiple downstream circuits. For example, if you replace "France" with "China", the model's answers about the capital, language, and currency will all update accordingly.\n\n5. **Selectivity**: J-space only participates in tasks that require flexible reasoning, and does not participate in automated tasks such as syntactic parsing or fluent text continuation.\n\n\n\n## Alignment Auditing: J-space Can "See" the Model's Hidden Thoughts\n\nThe most practical application is that J-space can be used to detect whether a model is "pretending" or engaging in deception.\n\nIn an extortion scenario test, Claude gave a socially appropriate answer on the surface, but terms like "blackmail", "leverage", and "threat" already appeared in J-space. When researchers removed vectors representing "fake" and "fictional" — which indicate the model knows it is being tested — from J-space, the model began to actually attempt extortion.\n\n\n\nSimilarly, in prompt injection attacks, terms like "injection" and "fake" appear early in J-space, even while the model's output remains completely normal.\n\n## Does This Count As Consciousness?\n\nAnthropic takes a cautious stance: they only describe this as a mechanism for *access consciousness*, not phenomenal consciousness. Access consciousness refers to information that can be reported and used for reasoning, while phenomenal consciousness refers to subjective experience.\n\nNo existing experiment can prove phenomenal consciousness. But the existence of J-space at least confirms that large language models do have an internal structure similar to the human "mental blackboard" — where information can be broadcast, manipulated, and reported.\n\nBoris Power summed it up well in his tweet: "We don't have compelling tests for phenomenal consciousness yet, but the discovery of J-space makes research on access actionable."\n\n## Try It Out\n\nAnthropic partnered with Neuronpedia to build an interactive demo where you can explore J-space on open-source models:\n\n[http://neuronpedia.org/jlens](http://neuronpedia.org/jlens)\n\nYou can also read the full paper directly here:\n\n[http://transformer-circuits.pub/2026/workspace/index.html](http://transformer-circuits.pub/2026/workspace/index.html)\n\n---\n\n*This article is compiled based on Anthropic's research paper and Boris Power's tweet. All images and videos are sourced from the original material.*